Download Computational Methods for Agricultural Research - wiki DPI

Survey
yes no Was this document useful for you?
   Thank you for your participation!

* Your assessment is very important for improving the workof artificial intelligence, which forms the content of this project

Document related concepts

Spatial analysis wikipedia , lookup

World Values Survey wikipedia , lookup

Map wikipedia , lookup

Stephen Grossberg wikipedia , lookup

Recurrent neural network wikipedia , lookup

Transcript
Computational Methods
for Agricultural Research:
Advances and Applications
Hercules Antonio do Prado
Brazilian Agricultural Research Corporation & Catholic University of
Brasilia, Brazil
Alfredo Jose Barreto Luiz
Brazilian Agricultural Research Corporation, Brazil
Homero Chaib Filho
Brazilian Agricultural Research Corporation, Brazil
InformatIon scIence reference
Hershey • New York
Director of Editorial Content:
Director of Book Publications:
Acquisitions Editor:
Development Editor:
Typesetter:
Production Editor:
Cover Design:
Kristin Klinger
Julia Mosemann
Lindsay Johnston
Joel Gamon
Deanna Jo Zombro
Jamie Snavely
Lisa Tosheff
Published in the United States of America by
Information Science Reference (an imprint of IGI Global)
701 E. Chocolate Avenue
Hershey PA 17033
Tel: 717-533-8845
Fax: 717-533-8661
E-mail: [email protected]
Web site: http://www.igi-global.com
Copyright © 2011 by IGI Global. All rights reserved. No part of this publication may be reproduced, stored or distributed in
any form or by any means, electronic or mechanical, including photocopying, without written permission from the publisher.
Product or company names used in this set are for identification purposes only. Inclusion of the names of the products or companies does not indicate a claim of ownership by IGI Global of the trademark or registered trademark.
Library of Congress Cataloging-in-Publication Data
Computational methods for agricultural research : advances and applications /
Hercules Antonio do Prado, Alfredo Jose Barreto Luiz, and Homero Chaib Filho,
editors.
p. cm.
Includes bibliographical references and index.
ISBN 978-1-61692-871-1 (hardcover) -- ISBN 978-1-61692-873-5 (ebook) 1.
Agriculture--Research--Data processing. 2. Agricultural informatics. I.
Prado, Hercules Antonio do. II. Luiz, Alfredo Jose Barreto, 1963- III. Filho,
Homero Chaib.
S540.D38C66 2010
630.72--dc22
2010035367
British Cataloguing in Publication Data
A Cataloguing in Publication record for this book is available from the British Library.
All work contributed to this book is new, previously-unpublished material. The views expressed in this book are those of the
authors, but not necessarily of the publisher.
107
Chapter 7
Using Self-Organizing Maps
for Rural Territorial Typology
Marcos Aurélio Santos da Silva
Brazilian Agricultural Research Corporation, Embrapa Coastal Tablelands, Brazil
Edmar Ramos de Siqueira
Brazilian Agricultural Research Corporation, Embrapa Coastal Tablelands, Brazil
Olívio Alberto Teixeira
Federal University of Sergipe, Department of Economy, Cidade Universitária, Brazil
Maria Geovania Lima Manos
Brazilian Agricultural Research Corporation, Embrapa Coastal Tablelands, Brazil
Antônio Miguel Vieira Monteiro
National Institute for Space Research, Image Processing Division, Brazil
AbSTRAcT
This work assessed the capacity of the self-organizing map, an unsupervised artificial neural network,
to aid the process of territorial design through visualization and clustering methods applied to a multivariate geospatial temporal dataset. The method was applied in the case study of Sergipe’s institutional
regional partition (Territories of Identity). Results have shown that the proposed method can improve the
exploratory spatial-temporal analysis capacity of policy makers that are interested in territorial typology. A new partition for rural planning was elaborated and confirmed the coherence of the Territories
of Identity.
INTRODUcTION
Rural territorial typology refers to the classification of commonalities found in contiguous rural
areas that may be useful for a well definition of
DOI: 10.4018/978-1-61692-871-1.ch007
public policies and to discover identity elements. In
an integrational perspective, all dimensions must
be considered in a spatial analysis. For this study,
it was preferred to adopt the definition of territory
used by the Ministry of Agrarian Development
(MDA, 2005, p.7).
Copyright © 2011, IGI Global. Copying or distributing in print or electronic forms without written permission of IGI Global is prohibited.
Using Self-Organizing Maps for Rural Territorial Typology
... territories are defined as a physical space,
geographically defined, generally continuous,
encompassing urban and rural, characterized by
multidimensional criteria, such as environment,
economy, society, culture, politics and institutions, and a population with relatively distinct
social groups that relate internally and externally
through specific processes, which can distinguish
one or more elements that indicate identity and
social, cultural and territorial cohesion.
Territories break down false boundaries (political divisions) among areas and facilitate the
process of socio-cultural identity construction,
discovery or investigation. The design of public
policies also benefits from this strategy, in a way
that watersheds, forests, ecosystems and other
homogeneous zones can be analyzed altogether.
Naturally, there are convergences and divergences
about territorial definition, mainly concerning the
focus of the study, for example, economical versus
environmental analysis. A regional development
study implies the use of an interdisciplinary approach, which reveals itself as a very difficult task
due to the complexity and uncertainty of this field.
Regional studies comprehend constraining factors,
such as scales, physical amplitude, quantity and
quality of data, diversity of applications, political
and practical aspects (Sabourin & Teixeira, 2002).
The general definition of territory by the MDA
(2005) argues that social-cultural-economical
cohesion is the main factor to establish a regional
aggregation into a territory. Although landscape
homogeneity and distribution of natural resources
are important aspects, they are not determinant.
Therefore, multivariate socio-economical geospatial data integration and interpretation can be
useful for policy makers that engage themselves
in the process of territory design.
In Brazil, there are many initiatives towards the
institutionalization of territories as a new strategy
for public management (MDA, 2005; Bandeira,
2006). In this context of interdisciplinary studies
and spatial integration, mainly for sustainable de-
108
velopment, rural and urban issues are not viewed
as different things, but treated as a systematic
problem (Sabourin & Teixeira, 2002; Flores,
2004; MDA, 2005). Besides, the new trend for
self-development by bottom-up as stated by Claval
(2008) increases the demand for new and practical
approaches to study and generate knowledge for
a good territorial partition.
The absence of a territorial theory (Abramovay, 2006) and the rise of territorial policies, at
least in Brazil, imply the demand for exploratory
methods that could solve practical problems, such
as exploratory spatial clustering and, at the same
time, it gives us helpful insights for a territorial
theory. There are many classical quantitative
strategies to perform a multivariate exploratory
analysis of geospatial data such as: colored maps,
multivariate statistics coupled with these maps,
spatial statistics, and geostatistics (Bailey, 1995).
In this study, an Artificial Neural Network (ANN)
was applied simultaneously for geospatial data
visualization and clustering. The objective was to
assess the capacity of the ANN, more precisely
the Self-Organizing Map (SOM), to find hidden
patterns in the geospatial dataset that could be
useful for a territorial typology.
This ANN was applied in the analysis of Sergipe’s institutional regional partition (Territories
of Identity) created by Teixeira et al. (2007) using both subjective and analytical methods. On
the other hand, the approach of this work used
quantitative geospatial dataset aggregated by
municipalities. Results were compared with this
official Sergipe’s clustering. These territories
were chosen because a great effort for Sergipe’s
territorial partition was performed by academics
and officials, so a huge amount of information
was gathered and organized. Consequently, the
know-how and knowledge about the way the study
of regional commonalities is performed increased
and, therefore, demanded for integrated quantitative analysis of multisource and multivariated
geospatial data. In general terms, the main goal is
to establish a semi-automatic territorial typology
Using Self-Organizing Maps for Rural Territorial Typology
process to aid the rural planning by identification
of homogeneous areas in Sergipe, Brazil.
MATERIAL AND METHODS
Self-Organizing Maps (SOM)
The Kohonen Self-Organizing Map (SOM) is a
competitive artificial neural network structured
in two layers (Kohonen, 2001). The first one
represents the input vector of the dataset variables xk=[ξ1, ξ2,…, ξd]T, k=1,2,…n, where n is
the number of input vectors and d corresponds to
data dimension, and the second one is a neuron’s
grid, usually bidimensional and fully connected
(Figure 1a). Each neuron j has one codevector
wj=[wj1,wj2,…,wjd]T associated to it. The main goal
of this ANN is to order, by a learning algorithm,
the input dataset into the neural grid, using the
codevectors (wj) as representations of the input
data (xk).
There are different topologies for the SelfOrganizing Map architecture, while the most
common is the two-dimensional topology. Figure
1 shows a one-dimensional SOM (b) and a toroidal SOM (c).
The learning process can be explained in three
phases. In the first phase (competitive) each xk
is presented to all neurons in search of the Best
Match Unit (BMU) using Euclidean distance as
a measurement reference of the distance feature.
In the second phase (cooperative) a neighborhood
relation between the BMU and the other neurons
is defined for a smoothed update of each wj. Finally, in the last phase (adaptive), all codevectors
wj will be updated using some kind of adaptive
rule, in general sequential or batch, see Equations
1 and 2 for the batch update rule used in this work
(Vesanto, 1999).
nV
i
si (t ) = ∑ x i
(1)
j
m
wi (t + 1) =
∑h
j
ji
(t )s j (t )
m
∑n
j
Vj
(2)
h ji (t )
Where si represents an input pattern sum for
the ith-Voronoi region, Vi; and nVi is the number of
samples for the Voronoi dataset of the ith-neuron.
The neighborhood kernel function between neu-
Figure 1. (a) A complete self-organizing NxM map architecture representation. (b) A one-dimensional
SOM. (c) A toroidal SOM
109
Using Self-Organizing Maps for Rural Territorial Typology
rons i and j at time t is represented by hij(t)=exp(dij2/2*δ(t)2), where δ(t)= δ(0)*exp(-t/C), dij is the
distance between neurons i and j on the neural grid,
C is a constant, and m is the number of Voronoi
regions (number of neurons). See Equation 3 for
the sequential update rule.
wji (t + 1) = wji (t) + α(t) hij (t) [xik (t) − wij (t)]
(3)
Where α(t), α(0)<1, represents the learning
rate function. The learning rate function α(t), hij(t)
and δ(t) are monotonically decreasing functions.
After the learning process, codevectors should
approximate, in a non-linear manner, the input
dataset. In addition, SOM preserves the topological structure of the input data, and nearby patterns
in the sample dataset are associated with nearby
neurons in the SOM grid. For this work, a bidimensional SOM, the batch learning algorithm
associated with hexagonal lattice, gaussian neighborhood kernel function, and linear initialization
for visualization were used. For the automatic
clustering it was used a uni-dimensional SOM.
The competitive process is the most timeconsuming of the learning process. Usually this
is a sequential search for the Best Match Unit
(BMU). This process can be optimized by using
some mechanism to minimize the heuristic search
or through the parallelization of the learning code
(Openshaw & Turton, 1996).
The parameters of learning are defined empirically, based on user experience and methods
of trial and error. However, some techniques
for automatically determining the parameters
of learning have been proposed, either through
genetic algorithms or numerical methods (Haese
& Goodhill, 2001).
The dimensionality of self-organizing map and
its size (m) depend on the type of problem and
purpose. The literature shows that determining the
size of the SOM is an empirical process (Flexer,
2001; Kohonen, 2001). In general, bi-dimensional
SOM is used because of its ability to project the
110
data of high dimension d into a two-dimensional
neural map grid.
Assessment of Quality of
SOM Learning Algorithm
There is a reasonable set of methods for assessing
the quality of the generated map after the learning
process. The most applied are the quantization
error vector and the topological error (Kohonen,
2001).
The quantization error (Eq), Equation 4, is the
average error of values of the difference between
the feature vector xk and wBMU codevector, that
is the winner in the competitive process for the
vector xk:
n
Eq =
∑x
k =1
k
− wBMU
n
(4)
The topological error (Et), Equation 5, evaluates how the structure of the neural grid approximate vectors of the input space. Whereas each
xk has a BMU as the first neuron in the order of
competition in the neural grid, BMU2 correspond
to second neuron at this scale. Thus, the error will
correspond to the percentage of BMU and BMU2
whose are not neighbors in the grid for the same xk:
Et =
1
n
n
∑ u(x
k =1
k
)
(5)
Where u(xk) is setted to one if the BMU and
BMU2 are not neighbors, and zero otherwise.
Data Visualization
After the learning process it is necessary to visually verify the result of topological ordering onto
the SOM’s grid. There are many ways of visual
representation of the dataset using SOM (Vesanto,
1997; Vesanto, 1999). For this study, two of the
Using Self-Organizing Maps for Rural Territorial Typology
most common methods were used. The first is the
direct link between the artificial neuron unit and
the xk, in this case, the municipality.
The second form, the Component Planes, uses
the values of each component of the codevector
wj to color the Self-Organizing Map. This method
allows the visual analysis of component correlation, outlier detection and spatial component
distribution on SOM (Vesanto, 1999; Kohonen,
2000; Silva et al., 2004a).
Two periods were used and such temporal aspect was included in the SOM model by mounting
these two datasets in only one, so each temporal
set of values for each municipality became an
entry xk. This is the simplest strategy, but yet
useful when the purpose in mind is a visual clustering and interpretation by component planes.
According to Barreto & Araújo (2001), there is
a plethora of SOM models for time processing,
but the most common application of the SOM in
sequence processing is the trajectory of BMU’s
in the latent map space (Varsta, 2002). Therefore,
these two periods were compared visually by
Component Planes following the trajectory of
each municipality onto the neural map. To proceed
this visual analysis, the last version of the SOM
ToolBox was used (www.cis.hut.fi/somtoolbox/).
Automatic clustering
The SOM’s codevectors can also be segmented
by a wide range of strategies, including the visual
ones (Vesanto & Alhoniemi, 2000). Costa (1999)
partitioned the dataset throughout the segmentation of the image generated by the codevector’s
matrix of distance. However, in general, this is
too complex and it does not exhibit clearly hidden
patterns. Vesanto (1999) used SOM as a data compressor for a posterior statistical analysis, but this
strategy needs a high level of human interaction.
Costa & Andrade Netto (2003) used a conceptually simple and empiric approach (Costa-Netto
algorithm) using only the information stored in
codevectors after the learning process as the activ-
ity level of neurons, the quantization error and the
neighborhood relation for an automatic clustering.
Considering the output layer as a non-oriented
graph structure, Costa & Andrade Netto (2003)
proposed a SOM clustering based on graph segmentation, in a way that the algorithm eliminates
inconsistent connections among neurons, in attempt to find the ideal partition of the SOM’s grid.
The algorithm is based on the geometric distance
information among neurons, quantization error
and level of the neuron’s activity. The strategy
is to use heuristic rules to eliminate inconsistent
connections between two neighbor neurons.
The algorithm:
1.1. Take all distances d between adjacent neurons
i and j using the codevector wj as a measurement parameter, d(wi,wj); and activity level
for each neuron i, H(i).
1.2. For each couple of adjacent neurons, i and
j, the edge will be considered inconsistent
if:
1.2.1.
the distance between codevectors
exceed by 2 the median distance of the
other adjacent neurons to i or j;
1.2.2.
two adjacent neurons have the
level of activity (H) less than 50% of
the allowed minimum (Hmin), or some
of them are inactive neurons (H(i) =
0); Hmin=ωHmed, 0.1 ≤ ω ≤ 0.6 and Hmed
= n/m;
1.2.3.
the distance between the centroids
of the dataset associated to neurons i
and j exceeds by 2 the distance between
the codevectors of i and j, d(wi, wj).
1.3. Remove all inconsistent edges.
1.4. Associate a unique code for each set of connected neurons.
At the end of the process, each group of connected neurons will represent one cluster of the
codevectors. The algorithm uses some empirical
values, defined by the user, but can segment the
input data using only the SOM information after the
111
Using Self-Organizing Maps for Rural Territorial Typology
learning process. To validate the clusters partition
of the dataset, it was used the Compose Density
between and within the clusters (CDbw) index
(Halkidi & Vazirgiannis, 2002; Silva et al., 2004b;
Wu & Chow, 2004). For this spatial analysis, the
TerraView software (www.dpi.inpe.br/terraview)
was used, as well as the TerraSOM pluging developed from the SOMCode library (www.cpatc.
embrapa.br/somcode/) that implements some
Self-Organizing Map related algorithms.
Dataset
The State of Sergipe has an area of 21,910.34
km2, accounting for 0.26% of national territory
and 1.4% of the Northeast. Its absolute position
is between the parallel 9° 31’ and 11º 34’ south
latitude and the meridians 36º 25’ and 38 º 14’
longitude West of Greenwich (Figure 2). Although
it constitutes a state of small size and population, Sergipe occupies a privileged position in
the economic and social development scenario
Figure 2. Municipalities of Sergipe’s state
112
of the Northeast, considering its recent entry
into modern economic activities: off-shore oil
exploitation, mining, tourism and ethanol production by sugar cane crops. The agriculture share
in the GDP corresponds to approximately 3,0%,
but it is very important in terms of employment
and food supply.
Sergipe’s economy is diversified. Nevertheless, it is important to highlight the mineral extractive-industry (based on non-metallic minerals)
located in cities within a radius of up to 40 km
from the capital, which intensely contributes to
the concentration of the population, income and
wealth in this area. Agricultural activities are
distributed throughout the whole territory, occupying a significant area, as well as labor force,
with a strong emphasis on family work. Sugar
cane and orange stand out, as well as cattle and
poultry. The tertiary sector has been showing
changes with the addition of new functions, especially in the service sector (Teixeira et al., 2007).
Using Self-Organizing Maps for Rural Territorial Typology
For this study, 44 indicators (Table 1) were
selected to represent demographic, social, economic, and agricultural production dimensions,
as well as agricultural activities, such as land use
and crop and animal production. Two periods (A
and B) were considered in this study for a comparison between the 1990’s and 2000’s values.
All variables were standardized according to their
meaning and standard deviation. These variables
were chosen according to their disponibility and
proximity with those used by Teixeira et al. (2007).
The typology process used the institutional
partition of the State of Sergipe held by the Department of Planning as a starting point, named
Territory of Identity. This process takes into account quantitative indicators (Table 1), popular
opinions gathered by workshops and multidisciplinary assessment of spatial aggregation (Teixeira et al., 2007). The main goal of this partition
was to find out a well-stablished division of the
State that could aid the process of general public
policies design. Figure 3 shows the eight Territories of Identity: Central Wasteland, High Hinterland, Low San Francisco, South Center, Great
Aracaju, East, Middle Hinterland and South.
Figure 3. Territories of identity
The High Hinterland Territory is formed by
seven municipalities: Canindé do São Francisco
(CSF), Grararu (GRR), Monte Alegre de Sergipe
(MAS), Nossa Senhora da Glória (GLR), Nossa
Senhora de Lourdes (LOU), Poço Redondo (PRD)
and Porto da Folha (PFL). This territory represents
22.3% of the states’s land area, covering 4908
km2. In 2007, the High Hinterland Territory had
a population of 137,926 inhabitants, with a density of 28.15 hab/Km2, representing 7.11% of the
total population of the state. In 2005, the GDP of
this Territory represented 11.12% of the state’s
GDP, with a GDP per capita of R$ 10,730.00. The
Index of Human Development in 2000 was 0.575.
The Low San Francisco Territory is formed
by fourteen municipalities: Amparo do São
Francisco (ASF), Brejo Grande (BRJ), Canhoba
(CNH), Cedro de São João (CED), Ilha das Flores
(IFL), Japoatã (JPT), Malhada dos Bois (MLH),
Muribeca (MUR), Neópolis (NEO), Pacatuba
(PAC), Propriá (PRO), Santana do São Francisco
(SSF), São Francisco (SFR) and Telha (TEL).
This territory corresponds to 9.0% of the state’s
land area, which represents 1986.3 km2 of total
area. In 2007, the Low San Francisco Territory
had a population of 123,482 inhabitants, with a
density of 63.45 hab/Km2, representing 6.37%
of the population of the state. In 2005, the GDP
of this Territory represented 3.94% of the state’s
GDP, with a GDP per capita of R$ 4,176.00. The
Index of Human Development in 2000 was 0.614.
The Middle Hinterland Territory is formed
by six municipalities: Aquidabã (AQB), Cumbe
(CUM), Feira Nova (FNV), Graccho Cardoso
(GCR), Itabi (ITB) and Nossa Senhora das Dores
(DOR). This territory represents 7.3% of the land
area of the state, corresponding to 1612.6 km2 of
total area. In 2007, the Middle Hinterland Territory had a population of 62,644 inhabitants, with
a density of 39.59 hab/Km2, representing 3.23%
of the population of the state. In 2005, the GDP
of this Territory represented 1.57% of the state’s
GDP, with a GDP per capita of R$ 3,298.00. The
Index of Human Development in 2000 was 0.621.
113
Using Self-Organizing Maps for Rural Territorial Typology
Table 1. List of analyzed variables
VARIABLE
1
POPTOT
DESCRIPTION
Total of resident population (people)
GROUP A
GROUP B
Pop. 1991
Pop. 20002
1
2
POPRUR
Pop. 1991
Pop. 2000
3
PRODMCOW
Milk cow production (1.000 l)
Census 19963
Census 20064
4
PRODMGOAT
Milk goat production (1.000 l)
Census 1996
Census 2006
5
PRODEGGS
Egg production (t)
Census 1996
Census 2006
6
TOTCOX
Cattle - Ox (unit)
Census 1996
Census 2006
7
TOTCS
Cattle - swine (unit)
Census 1996
Census 2006
Birds (unit)
Census 1996
Census 2006
Tractors (amount)
Census 1996
Census 2006
Total of rural resident population (people)
8
TOTBIRDS
9
TRACTORS
10
PAREAPERM
Permanent and temporary crops (ha)
Census 1996
Census 2006
11
PAREAPAST
Natural and artificial pastures (ha)
Census 1996
Census 2006
12
PAREAFOT
Cultivated and natural forests (ha)
Census 1996
Census 2006
13
PMCFSJ
Percentage of women heads of households with no spouse and
with minor children (15 years old)
ATLAS 19915
ATLAS 20005
14
PPOOR
Percentage of poor people
ATLAS 1991
ATLAS 2000
15
PCI
Percentage of indigent children
ATLAS 1991
ATLAS 2000
16
PC714FE
Percentage of children out of school (range from 7 to 14 years
old)
ATLAS 1991
ATLAS 2000
17
INPERCAPITA
Per capita income (R$)
ATLAS 1991
ATLAS 2000
Income’s percentage related to governmental transferences
ATLAS 1991
ATLAS 2000
Income’s percentage related to work’s yield
ATLAS 1991
ATLAS 2000
Relation between 10% richest and 40% poorest
ATLAS 1991
ATLAS 2000
18
PINGOV
19
PINWORK
20
RBRICSPOOR
21
GINI
Gini’s index
ATLAS 1991
ATLAS 2000
22
ITHEIL
Theil’s index
ATLAS 1991
ATLAS 2000
23
INTIND
Indigence intensity
ATLAS 1991
ATLAS 2000
24
IDHMIN
Municipal Human Development Index - Income
ATLAS 1991
ATLAS 2000
25
IDHMLONG
Municipal Human Development Index - Longevity
ATLAS 1991
ATLAS 2000
Life expectancy at birth
ATLAS 1991
ATLAS 2000
Municipal Human Development Index - Education
ATLAS 1991
ATLAS 2000
26
ESPLIFE
27
IDHMEDUC
28
RILLI
Rate of illiteracy
ATLAS 1991
ATLAS 2000
29
GRSA
Gross rate of school attendance
ATLAS 1991
ATLAS 2000
30
RFECUN
Rate of fecundity
ATLAS 1991
ATLAS 2000
31
PPDOWN
Percentage of people living in their own and paid off house or
land
ATLAS 1991
ATLAS 2000
32
P714A
Percentage of illiterate children (ranging from 7 to 14 years old)
ATLAS 1991
ATLAS 2000
33
PPWATER
Percentage of people living in houses with piped water
ATLAS 1991
ATLAS 2000
34
PPWC
Percentage of people living in houses with WC and piped water
ATLAS 1991
ATLAS 2000
35
PPGARB
Percentage of people living in houses with garbage service
ATLAS 1991
ATLAS 2000
36
PPENERGY
Percentage of people living in houses with electric energy
ATLAS 1991
ATLAS 2000
37
PPIBAGRI
IPEA 19966
IPEA 20066
Percentage of Agricultural GDP
continued on following page
114
Using Self-Organizing Maps for Rural Territorial Typology
Table 1. continued
VARIABLE
38
RTTotal
39
MPRODMAIZE
DESCRIPTION
Constitutional transferences from National Treasure
GROUP A
GROUP B
MF 19967
MF 20067
Maize production (t)
8
Mean 90/96
Mean 00/068
Cassava production (t)
Mean 90/96
Mean 00/06
Bean production (t)
Mean 90/96
Mean 00/06
Sugar cane production (t)
Mean 90/96
Mean 00/06
40
MPRODCAS
41
MPRODBEAN
42
MPRODSC
43
MPRODORAN
Orange production (t)
Mean 90/96
Mean 00/06
44
MPRODCOCO
Coconut production (1.000 fruits)
Mean 90/96
Mean 00/06
SOURCE: Demographic census 1991 - IBGE. Demographic census 2000 - IBGE. Agricultural census 1995/1996 - IBGE. 4Agricultural census 2006 - IBGE. 5Atlas of Human Development - PNUD/Brasil. 6IPEADATA (http://www.ipeadata.gov.br/). 7Ministry of Finance National Treasure (http://www.tesouro.fazenda.gov.br/estados_municipios/transferencias_constitucionais.asp). 8Anual Municipal Agricultural Production - PMA/IBGE.
1
2
The East Territory is formed by nine municipalities: Capela (CPL), Carmópolis (CAR),
Divina Pastora (DVN), General Maynard (GMN),
Japaratuba (JTB), Pirambu (PIR), Rosário do Catete (RCT), Santa Rosa de Lima (SRL) and Siriri
(SRR). This territory corresponds to 6.7% of the
land area of the state, possessing 1474.1 km2 of
total area. In 2007, the territory of East Sergipe
had a population of 90,452 inhabitants, with a
density of 59.56 hab/Km2, representing 4.66%
of the population of the state. In 2005, the GDP
of this Territory represented 6.47% of the state’s
GDP, with a GDP per capita of R$ 9,878.00. The
Index of Human Development in 2000 was 0.643.
The Central Wasteland Territory is formed
by fourteen municipalities: Areia Branca (ARB),
Campo do Brito (CPB), Carira (CRR), Frei
Paulo (FPA), Itabaiana (ITA), Macambira (MAC),
Malhador (MLD), Moita Bonita (MOI), Nossa
Senhora Aparecida (APA), Pedra Mole (PRM),
Pinhão (PIN), Ribeirópolis (RIB), São Domingos
(SDO) and São Miguel do Aleixo (SMA). This
territory corresponds to 6.7% of the land area of
the state, possessing 1474.1 km2 of total area. In
2007, the Central Wasteland Territory Territory
had a population of 222,197 inhabitants, with a
density of 71.14 hab/Km2, representing 11.46%
of the population of the state. In 2005, the GDP
of this Territory represented 6.70% of the state’s
3
GDP, with a GDP per capita of R$ 3,949.00. The
Index of Human Development in 2000 was 0.646.
The South Territory is formed by eleven
municipalities: Arauá (ARA), Boquim (BOQ),
Cristinápolis (CRI), Estância (EST), Indiaroba
(IND), Itabaianinha (ITN), Pedrinhas (PDR),
Salgado (SAL), Santa Luzia do Itanhy (SLI),
Tomar do Geru (TGE) and Umbaúba (UMB).
This territory represents 14.5% of the land area of
the state, with 3193.6 km2 of total area. In 2007,
the South Territory had a population of 241,192
inhabitants, with a density of 77.06 hab/Km2, representing 12.44% of the population of the state. In
2005, the GDP of this Territory represented 8.40%
of the state’s GDP, with a GDP per capita of R$
4,579.00. The Index of Human Development in
2000 was 0.616.
The South Center Territory is formed by
five municipalities: Lagarto (LGT), Poço Verde
(PVD), Riachão do Dantas (RDT), Simão Dias
(SDI) and Tobias Barreto (TBA). This territory
represents 16.1% of the land area of the state,
possessing 3551.5 km2 of total area. In 2007,
the South Center Territory had a population of
213,492 inhabitants, with a density of 60.63 hab/
Km2, representing 16.07% of the population of
the state. In 2005, the GDP of this Territory represented 5.67% of the state’s GDP, with a GDP
per capita of R$ 3,480.00. The Index of Human
Development in 2000 was 0.599.
115
Using Self-Organizing Maps for Rural Territorial Typology
The Great Aracaju Territory is formed by
nine municipalities: Aracaju (AJU), Barra dos
Coqueiros (BCQ), Itaporanga d’Ajuda (IAJ),
Laranjeiras (LRJ), Maruim (MRI), Nossa Senhora
do Socorro (SOC), Riachuelo (RIA), Santo Amaro
das Brotas (SAB) and São Cristóvão (SCR).
This territory represents 9.9% of the land area of
the state, with 2192 km2 of total area. In 2007,
the Great Aracaju Territory had a population of
847,941 inhabitants, with a density of 387.65 hab/
Km2, representing 43.72% of the population of
the state. In 2005, the GDP of this Territory represented 56.12% of the state’s GDP, with a GDP
per capita of R$ 8,785.00. The Index of Human
Development in 2000 was 0.750.
GEOSPATIAL ANALYSIS bY
SELF-ORGANIZING MAPS
The analysis of multivariate geospatial data could
be pointed as one of the great scientific challenges
to be solved in the present. Several institutions
are providing amounts of geo-referenced data
over the Internet, but there are many limitations
to the effective use and interpretation of that type
of dataset. In fact, the integration of all information depends on the ability to analyze the data
and find relevant knowledge in such databases.
As stated by Bação et al. (2005) “More than
prediction tools, we need to develop exploratory
tools based on classification and clustering, which
would enable an improved understanding of the
available data” (p. 155).
The Self-Organizing Map has been applied
to the problem of geospatial analysis in many
ways. Nevertheless, in general, the SOM is used
for exploratory spatial analysis of census data
aggregated by area. The way the analysis is performed varies, but the use of visual interpretation
methods is registered in almost every application.
Winter & Hewitson (1994) applied the SOM
associated with the Component Planes for the
characterization of 21 districts of the Western Cape
116
of South Africa using 10 census variables. From
the Component Planes it was possible to characterize mainly the black population of the region.
Kaski & Kohonen (1996) showed the feasibility
of the visualization of multivariate data from the
SOM. The dataset consisted of 39 indicators,
which described different aspects of the welfare
“states” of the countries around the world.
Koua (2003) analyzed 29 socioeconomic indicators related to municipalities. The author used
the Component Planes and the matrix of distances
among codevectors wj for a visual geospatial exploration and discovered that SOM improved the
geographical analysis and offered support for the
exploration of large geospatial datasets, increasing
knowledge in this area. Silva (2004ab) performed
a complete exploratory spatial analysis using eight
socio-economic indicators related to census sectors. The author adjusted the algorithms related
to the SOM neural network in a suite for outlier
detection, spatial dependence analysis, automatic
clustering and component analysis. The suite was
applied in the problem of intra-urban analysis of
social exclusion and was compared with statistical
analysis (Genovez, 2002).
Aksoy (2006) applied the SOM to the problem
of the regionalization of 923 districts in Turkey
using 36 socio-economic indicators and compared
the results with a statistical approach (k-means
and Principal Component Analysis). Spielman
& Thill (2008) studied the spatial relationship
among 2,217 census tracts in New York City using a dataset of 79 attributes applying a simple
method of geospatial data visualization linking
the artificial neural network with a Geographical
Information System (GIS).
There are some differences between the approach of this chapter and the others listed in this
short review. First, in the present study, there is an
explicit focus on rural aspects of territorial study
while the majority of researches focus on urban
problems; second, the chapter proposed a complete
regional characterization (typology) using, at the
same time, visual tools and an automatic clustering
Using Self-Organizing Maps for Rural Territorial Typology
algorithm; finally, the work started from a previous
qualitative territorial study, Territories of Identity,
that makes a comparison of the results possible.
RESULTS AND DIcUSSION
General Spatial Dependence
by Visual Interpretation
After SOM’s learning algorithm processing had
been performed, the correspondence between
each input vector (municipality-group) to its
corresponding artificial neuron was realized, according to the Euclidean distance. Thus, it was
possible to visualize the position in the neural
map of each municipality-group and examine the
proximity relations in the feature’s space among
them. Whereas the study has the institutional
division of the state of Sergipe into Territories
of Identity as its starting point, the neural map
was partitioned having it as guide, which means
looking for proximities among municipalities of
the same Territory of Identity.
From the visual analysis of the distribution
of the cities on the neural map it was found that
there is spatial dependence for most of the input
vectors. That is, neighboring municipalities in
geophysical space are also deeply related in the
space of attributes (neural map). From this observation it was possible to trace thresholds among
the regions of the neural map that corresponded
to most municipalities-group of each area previously defined (Territories of Identity). Therefore,
Figure 4a illustrates how the Territories of Identity are distributed in the neural map. It was also
observed that there is spatial dependence among
these territories.
Figure 4b shows the distribution of some
municipalities in the neural map so that it is possible to analyze two factors: discrepancies between
the positions of each municipality according to
data from two periods (A and B), and the interterritorial shifting of municipalities in the space
of attributes, so that it is possible to analyze the
link between the territories constructed by the
neural map and the Territories of Identity. It is
important to note that when significant differences between the groups A and B for the same
city are not present, it will be only represented by
the symbol of the city. Otherwise, the municipal-
Figure 4. (a) SOM after the visual partition. (b) Linking between each artificial neuron and the municipality associated to it. The number of municipalities represented onto the neural map was limited for
legibility reasons
117
Using Self-Organizing Maps for Rural Territorial Typology
ity will be represented by its initials followed by
a letter indicating the group (A or B).
From Figure 4b, it is possible to identify the
local displacement between periods A and B. The
municipality of Capela (CPL) migrated from the
region linked to the South Central region to the
Lower San Francisco Territory, approaching the
municipalities of the East. The municipality of
Poço Verde (PVD) migrated from the area linked
to the High Hinterland to the area linked to the
South Center. The municipality of São Miguel do
Aleixo (SMA) migrated from the region associated
with the Central Wasteland to the Middle Hinterland. The municipalities of Umbaúba (UMB)
and Boquim (BOQ) migrated from the Central
Wasteland to the South region. The municipality
of Itaporanga d’Ajuda (IAJ) migrated from the
South region to the region nearby the Central
Wasteland, into the direction of the Great Aracaju.
The municipalities of Indiaroba (IND), Pedrinhas
(PDR) and Arauá (ARA) scattered from the area
linked to the South, contrary to the behavior of
BOQ and UMB.
The analysis of temporal shifts between groups
A and B did not show significant changes for
most municipalities. The Southern region showed
greater variation between the two periods in the
sense of closeness between municipalities (BOQ
and UMB) and distance (IND, PDR, NIV). The
municipalities of CPL and PVD had the highest
displacements. The municipality of São Francisco
(SFR) migrated from the Low San Francisco to
the region associated to the Middle Hinterland.
Only half the regions (that represent the territories) formed by the neural map have at least
one of its Centralities1 not positioned in the
region of concentration of the majority of the
municipalities of the same Territory of Identity.
The municipality of Estancia (EST), for example,
belongs to the Southern Territory but is extremely
close to the Great Aracaju Territory on the neural
map. The municipalities of Aquidabã (AQB) and
Nossa Senhora das Dores (DOR) are far away
in the feature space of the other municipalities
118
of the Middle Hinterland Territory. Canindé do
São Francisco (CSF) is a centrality with characteristics similar to those of most municipalities
in the High Hinterland Territory, which does not
occur with the municipality of Nossa Senhora
da Glória (GLR), that stands in the region of the
South Central Territory. Another centrality that is
positioned distantly from the other municipalities
of the same Territory in the feature space is Propriá
(PRO), which also has similar characteristics to
the municipalities of the Great Aracaju. All the
centralities are shown in Figure 3b.
The same occurs with other municipalities that
are not Centralities. The municipality of Riachão
do Dantas (RDT) belongs to the South Central
Territory, but has features related to the South
Territory. The municipality of Nossa Senhora de
Lourdes (LOU) belongs to the High Hinterland
Territory, but has characteristics more intimately
related to the Middle Hinterland Territory. The
municipalities of Santo Amaro das Brotas (SAB),
Maruim (MRI), Riachuelo (RIA) and Laranjeiras
(LRJ) are associated to the Great Aracaju Territory,
in spite of presenting characteristics associated
with the municipalities of the East Territory.
It was observed that there are some differences
between the groups formed by the neural map and
the Territories of Identity, although, in general,
the neural approach confirms the cohesion intraTerritories in the feature space.
characterization by
component Planes
Through Components Planes (Figure 5) it was
possible to characterize each group defined by
neural map (that represents each territory) through
the observation of the spatial distribution of each
component intensity in the neural map. As each
municipality is associated with only one artificial
neuron it is possible to use the partition of the
neural map to facilitate the characterization of
the territories.
Using Self-Organizing Maps for Rural Territorial Typology
Figure 5. Component planes for each variable
First, it was observed that there were outliers
for some variables. Aracaju (AJU) was differentiated from the others with high values for the
variables POPTOT, INPERCAPITA, ITHEIL,
IDHMIN and RTTOTAL. The municipality of
Lagarto (LGT) has high scores for the variables
POPRUR and MPRODCAS. The High Hinterland
and South Center territories presents high values
for PRODMCOW, PRODMGOAT, TOTOCOX,
TOTBIRDS, TOTCS, MPRODMAIZE and
MPRODBEAN. Outliers were also detected for
PRODEGGS, PAREAPAST, PAREAFOT,
RBRICSPOOR variables. Correlation was also
observed, same patterns for the Component Plane
coloring, between variables PPOOR and PCI,
IDHMLONG and ESPLIFE, IDHMEDUC and
RILLI and PPWATER and PPWC.
The area which borders the High Hinterland
region on the neural map was characterized by
high values for the variables TOTCOX, PPOOR,
PCI, INTIND, PPDOWN, P714A, PPGARB,
PPIBAGRI and MPRODSC, average values for
the variables PC714FE and GINI, and low values
for the variables PAREAPERM, PMCFSJ, PINGOV, IDHMLONG, ESPLIFE, IDHMEDUC,
RILLI, GRSA, PPWATER, PPWC, PPENERGY,
MPRODORAN and MPRODCOCO.
The area which borders the Lower San Francisco region was characterized by high values for the
variables PPOOR, PCI, PINGOV, RFECUN and
PPENERGY, average values for the variables PMCFSJ, PINWORK, GINI, INTIND, IDHMEDUC,
RILLI, GRSA, PPDOWN, P714A, PPWATER,
PPWC and PPIBAGRI, and low values for the
variables TOTCOX, TOTBIRDS, TRACTORS,
PAREAPERM, PC714FE, IDHMLONG, ESPLIFE, PPGARB, MPRODSC, MPRODORAN
and MPRODCOCO.
The area which borders the Middle hinterland
region was characterized by high values for the
variables PPOOR, PCI, PINWORK, PPDOWN,
PPGARB and PPIBAGRI, average values for the
variables PINGOV, GRSA, RFECUN, P714A
and PPENERGY, and low values for the variables TOTCOX, TRACTORS, PAREAPERM,
PMCFSJ, PC714FE, GINI, INTIND, IDHMLONG, ESPLIFE, IDHMEDUC, RILLI, PPWATER, PPWC, MPRODSC, MPRODORAN and
MPRODCOCO.
The area which borders the East region on
neural map was characterized by high values for
119
Using Self-Organizing Maps for Rural Territorial Typology
the variables PMCFSJ, PPOOR, PCI, PPGARB,
PPENERGY and MPRODSC, average values
for the variables IDHMLONG, ESPLIFE, IDHMEDUC, RILLI, GRSA, RFECUN, PPDOWN,
PPWATER and PPWC and low values for the variables TOTCOX, TRACTORS, PAREAPERM,
PC714FE, P714A, PPIBAGRI, MPRODORAN
and MPRODCOCO.
The area which borders the Central Wasteland
region was characterized by high values for the
variables IDHMLONG, ESPLIFE, PPDOWN,
PPGARB and PPENERGY, average values for
the variables PPOOR, PCI, PC714FE, PINGOV,
PINWORK and PPIBAGRI, and low values for the
variables TOTCOX, TRACTORS, PAREAPERM,
PMCFSJ, GINI, INTIND, IDHMEDUC, RILLI,
GRSA, RFECUN, P714A, PPWATER, PPWC,
MPRODORAN and MPRODCOCO.
The area which borders the South region on the
neural map was characterized by high values for
the variables PPOOR, PCI, PC714FE, PINWORK,
PPDOWN, P714A, PPGARB, PPIBAGRI,
MPRODORAN, average values for the variables
TRACTORS, PMCFSJ, GINI, RFECUN, and low
values for the variables TOTCOX, PAREAPERM,
PINGOV, INTIND, IDHMLONG, ESPLIFE,
IDHMEDUC, RILLI, GRSA, PPWATER, PPWC,
PPENERGY, and MPRODCOCO.
The area which borders the South Central
region was characterized by high values for the
variables TOTCOX, PPDOWN, and PPGARB,
average values for the variables TRACTORS,
PPOOR, PCI, PC714FE, PINGOV, PINWORK,
GINI, INTIND, RFECUN, P714A, PPWATER,
PPWC, PPENERGY and PPIBAGRI, and low
values for the variables PAREAPERM, 13.IDHMLONG, ESPLIFE, IDHMEDUC, RILLI, GRSA,
MPRODORAN and MPRODCOCO.
The area which borders the Great Aracaju region on neural map is characterized by high values
for the variables PINGOV, GINI, IDHMEDUC,
RILLI, GRSA, PPWATER, PPWC, PPGARB
and PPENERGY, average values for the variables
PAREAPERM, PMCFSJ, PC714FE, PINWORK,
120
INTIND, IDHMLONG, ESPLIFE and MPRODCOCO and low values for variables TOTCOX,
TOTBIRDS, TRACTORS, PPOOR, PCI, RFECUN, PPDOWN, P714A and PPIBAGRI.
It was observed that the municipalities of the
High Hinterland region form a more homogeneous, cohesive and well defined region, while
the Central Wasteland Territory, as well as the
Low San Francisco, have some variation degree
which could imply the existence of subdivisions.
This suggests that policy makers should pay
attention to internal territorial differences and
particularities and decide according to the best
solution of each subdivision. Other neural map
territories have shown good degree of cohesion
by the Component Planes analysis.
Automatic clustering by TerraSOM
From the distribution of the municipalities on
the neural map and the variables analysis by the
Component Planes it was possible to propose a
new arrangement of the Territories of Identity.
Although the neural map regionalization by visual
interpretation was not conclusive, it was possible to
proceed with a third clustering by the Costa-Netto
algorithm to improve the amount of information
about the aggregation of Sergipe’s municipalities
and validate the previous results.
Figure 6 shows the regionalization of the state
of Sergipe by a one-dimensional SOM, with 10
neurons and linearly initiated. Several network
sizes were tested, but basically the same groups
were found, varying not more than in terms of
aggregation or subdivision of each one. The automatic clustering algorithm found 6 groups, as
shown in Figure 6. Important to note that there
was no contiguity restriction on the clustering
algorithm, so non-neighboring municipalities
clusters were allowed.
The automatic regionalization confirms the
previous results. The municipality of Riachão do
Dantas belongs to the group 5, the same group of
municipalities in the South Territory (Tomar do
Using Self-Organizing Maps for Rural Territorial Typology
Figure 6. Automatic regionalization by the TerraSOM plugin without neighborhood constraints
Geru, Indiaroba, Santa Luzia do Itanhy and Cristinápolis). The municipality of Estancia belongs
to the group 1, the same group of the municipalities of the Great Aracaju Territory. Also note
the distinction between these municipalities in
the groups 2 and 3, which belong to the Grand
Aracaju Territory. Nevertheless, according to the
SOM network, they belong to the East Territory
(Laranjeiras, Riachuelo, Maruim and Santo
Amaro das Brotas).
It was also confirmed by the automatic regionalization that the municipalities of Nossa Senhora
de Lourdes and São Miguel do Aleixo were closer
in the feature space with the Middle Hinterland
Territory. It was observed that the group 4 congregated municipalities from three Territories of
Identity (Lower San Francisco, Central wasteland
and Middle Hinterland), which demonstrates the
difficulty of separating these municipalities.
It was taken into account the contiguity among
municipalities during the process of construction
of the new partition. The consistency of the results
of the three forms of the SOM analysis improved
the consistency of the proposal.
Figure 7 shows the final configuration of the
state of Sergipe partition based on the forty-four
variables. The municipality of Riachão do Dantas
migrated from the South Central Territory to the
South Territory, while Estância migrated from the
Southern Territory to the Great Aracaju Territory.
The municipalities of Laranjeiras, Maruim,
Santo Amaro das Brotas and Riachuelo migrated
from the Great Aracaju Territory to the Eastern
Territory. The municipality of São Miguel do
Aleixo migrated from the Central Wasteland Territory to the Middle Hinterland Territory. And
finally, the city of Nossa Senhora de Lourdes
migrated from the High Hinterland Territory to
the Middle Hinterland Territory.
This final regionalization proposal is not
intended to replace the official one, but rather
demonstrates the ability to undertake the territorialization from an exploratory spatial analysis by
a neural network. Teixeira et al. (2007) evidenced
Figure 7. Proposed regionalization by SOM
algorithm
The Final Regional Partition
After the visual and automatic clustering and interpretation of the neural map a new regionalization
of the state of Sergipe was proposed (Figure 7).
121
Using Self-Organizing Maps for Rural Territorial Typology
that several key criteria to describe the association
to one or another territory is also intimately related
to group and personal abstract values, e.g., Nossa
Senhora de Lourdes (LOU) and São Miguel do
Aleixo (SMA) are quantitatively within the reach
of the Middle Hinterland, but the participative
process resulted in a different association between
these municipalities and the Territories of Identity.
Estância (EST), Riachão do Dantas (RDT) and
Itaporanga D’Ajuda (IAJ) are vast municipalities,
and the existence of concentrated activities could
lead to misclassification. However, considering
only the forty-four variables mentioned in this
study, it was concluded that Estância is more
related to the Great Aracaju Territory, once this
municipality has a great industrial GDP and a low
agricultural production. The geographical space
of the RDT municipality goes from the Wasteland
to the Hinterland, despite of the indication of a
deeper relation to the South Territory, especially to
the municipalities of the extreme south, as it can
be observed in Figure 5. Maruim (MRI), Laranjeiras (LRJ), Santo Amaro das Brotas (SAB) and
Riachuelo (RIA) present features nearer the East
Territory than the Great Aracaju. The sugar cane
activity, the governmental transferences (mainly
the royaties), and the way of social organization
may approximate this group to municipalities of
the East Territory.
FUTURE RESEARcH DIREcTIONS
The SOM acts as an extractor of features without reducing the size of the feature vector, i.e.,
without significant loss of statistical information on geospatial dataset. This allows the use
of statistical methods associated with the neural
network, especially in the post-processing of the
codevectors w. Hierarchical clustering methods
or algorithms based on vector quantization can
be applied to the clustering process of the codevectors. Depending on the relationship between
122
the number of input vectors and the size of the
feature vector, the application of the Principal
Component Analysis (PCA) could be necessary
for a dimensionality reduction.
There is also the possibility to add new components in the input dataset so that new factors could
be considered in the process of regionalization.
It is possible, for example, to include an explicit
component of spatial location of the municipalities
into the input vector or into the learning process
of the SOM algorithm for a spatial constraint.
Numerical and categorical data fusion can be
performed for a better representation of different
dimensions, both environmental and economical.
Additional temporal information can be added
for a trend analysis using the SOM’s projection
algorithms.
As an interdisciplinary study a lot of work is
necessary to join distinct concepts and actions from
many disciplines, such as economy, sociology,
geography and agronomy towards rural territorial
typology for a sustainable development. However,
this integrative and systemic approach is still a
challenger for researchers and policy makers.
cONcLUSION
The study of the sustainable development of rural
areas depends on the growth of integrated and
multidisciplinary studies. The regional approach
has been one of the methods to design projects
of development and regional research in order
to consider social, economic, geophysical and
cultural dimensions. Population growth, and
hence the demand for quality public services and
projects for regional infrastructure, require from
public managers greater flexibility in decisionmaking, so that they observe the requirements of
fairness, efficiency and effectiveness. Territorial
typology certainly provides important feedback
for the development of a collective solution to
issues relating to rural development.
Using Self-Organizing Maps for Rural Territorial Typology
The large volume of data already available
and their differences in origin and format imposes
research challenges, requiring new methods of data
analysis, especially those related to data mining.
The possibility of visualization of multivariate
data combined with the ability to use automatic
clustering algorithms turned the neural network
SOM into an important tool in the process of territorial typology.
Considering forty-four heterogeneous variables the Territory of Identity presented by the
dataset distribution on artificial neural map intra
and inter spatial dependence and low level of interterritorial shifting of municipalities (only 12%),
however, the analysis by Component Planes have
shown some subdivision intra-territories showing
that policy makers must be careful about simple
and homogeneous solutions for each Territory. The
automatic regionalization by the SOM algorithm
integrated with the Costa-Netto partition strategy
associated with the CDbw validity cluster index
confirmed that there are significant differences
intra-territories and that there are no-neighbor
municipalities with similar characteristics (for
example, Riachão do Dantas, Canindé do São
Francisco and Brejo Grande). The automatic
clustering has also shown that some Centralities
(Itabaliana, Estância and Propriá) are closer to the
Great Aracaju Territory and that some municipalities of the Great Aracaju Territory could be associated to another group or to the East Territory.
Based on the Territories of Identity and on the
results of the SOM algorithm was constructed a
regionalization for rural studies purposes. It can
help the general understanding about the Sergipes’s rural actual scenario and aid the process of
regional planning. This is particularly important
for regions where agriculture is a major employer
of labor, family labor and, therefore, important
source of income for the farming populations.
REFERENcES
Abramovay, R. (2006). Para una teoría de los estudios territoriales. In Manzanal, M., Neiman, G.,
& Lattuada, M. (Eds.), Desarrollo rural – Organizaciones, instituciones y territorios (pp. 51–70).
Buenos Aires, Argentina: Ediciones Ciccus.
Aksoy, E. (2006, October) Clustering with GIS:
an attempt to classify turkish district data. Paper
presented at the XXIII FIG Congress. Munich,
Germany.
Bação, F., Lobo, V., & Painho, M. (2005). The selforganizing map. the Geo-SOM, and relevant variants for geosciences. Computers & Geosciences,
31, 155–163. doi:10.1016/j.cageo.2004.06.013
Bailey, T. C., & Gatrell, A. C. (1995). Interactive
Spatial Data Analysis. Longman.
Bandeira, P. S. (2006). Institucionalização de
regiões no Brasil. Revista da Sociedade Brasileira
para o Progresso da Ciência, 58(1), 34-37.
Barreto, G. A., & Araújo, A. F. R. (2001). Time
in self-organizing maps: An overview of models.
International Journal of Computer Research,
10(2), 139–179.
Claval, P. (2008). Espaces et territoire: les bifurcations de la science régionale. Géographie,
Économie. Société, 10, 157–184.
Costa, J. A. F. (1999) Classificação automática
e análise de dados por redes neurais autoorganizáveis. Unpublished doctoral dissertation,
University of Campinas, Campinas, São Paulo.
Costa, J. A. F., & Andrade Netto, M. L. (2003).
Segmentação do SOM Baseada em Particionamento de Grafos. Paper presented at the VI Brazilian Congress on Neural Networks, São Paulo,
(pp. 451-456).
123
Using Self-Organizing Maps for Rural Territorial Typology
Flexer, A. (2001). On the use of self-organizing
maps for clustering and visualization. Intelligent
Data Analysis, 5, 373–384.
Flores, M. (2004). Desenvolvimento territorial
rural: uma proposta de estudo para apoio à formulação de políticas públicas. In Lages, et al (Ed.),
Territórios em movimento: cultura e identidade
como estratégia de inserção competitiva. Brasília,
DF: SEBRAE.
Genovez, P. (2002). Território e Desigualdades:
análise espacial intra-urbana no estudo da
dinâmica de exclusão/inclusão social no espaço
urbano em São José dos Campos-SP. Unpublished
master thesis, National Institute for Space Research, São José dos Campos, São Paulo, Brazil.
Haese, K., & Goodhill, G. J. (2001).Auto-SOM: Recursive Parameter Estimation for Guidance of SelfOrganizing Feature Maps. Neural Computation,
13, 595–619. doi:10.1162/089976601300014475
Halkidi, M., & Vazirgiannis, M. (2002). Clustering
validity assessment using multi representatives.
Paper presented at the SETN Conference, Thessaloniki, Greece.
Kaski, S., & Kohonen, T. (1996). Exploratory data
analysis by the self-organizing map: structures of
welfare and poverty in the world. In Refenes, A.-P.
N., Abu-Mostafa, Y., Moody, J., & Weigend, A.
(Ed.), Third International Conference on Neural
Networks in the Capital Markets (pp. 498-507).
Singapore: World Scientific.
Kohonen, T. (2001). Self-organizing maps. Berlin:
Springer.
Koua, E. L. (2003). Using self-organizing maps
for information visualization and knowledge
discovery in complex geospatial datasets. Paper
presented at the XXI International Cartographic
Conference (ICC), (pp. 1694-1702).
124
Ministério do Desenvolvimento Agrário – MDA.
(2005). Referências para a gestão social de territórios rurais. Brasília, DF: MDA. Série Documentos nº 3.
Openshaw, S., & Turton, I. (1996). A parallel
Kohonen algorithm for the classification of large
spatial datasets. Computers & Geosciences, 22(9),
1019–1026. doi:10.1016/S0098-3004(96)000404
Sabourin, E., & Teixeira, O. A. (2002). Planejamento e desenvolvimento dos territórios rurais.
Brasília: Embrapa Informação Tecnológica.
Silva, M. A. S., Monteiro, A. M. V., & de Medeiros,
J. S. (2004a). Visualization of geospatial data by
component planes and u-matrix. Paper presented
at the VI Brazilian Symposium on GeoInformatics,
Campos do Jordão, Brazil.
Silva, M. A. S., Monteiro, A. M. V., & de Medeiros,
J. S. (2004b). Semi-Automatic Geospatial Data
Clustering by Self-Organizing Maps. Paper presented at the VI Brazilian Symposium on Neural
Networks, 2004, São Luis, MA.
Spielman, S. E., & Thill, J.-C. (2008). Social
area analysis, data mining, and GIS. Computers,
Environment and Urban Systems, 32, 110–122.
doi:10.1016/j.compenvurbsys.2007.11.004
Teixeira, O. A., Melo, R. L. de & França, V. L.
A. (2007). Estado de Sergipe: uma proposta de
territorialização para o planejamento. SEPLAN
Report.
Varsta, M. (2002). Self-organizing maps in
sequence processing. Unpublished doctoral dissertation, Helsinki University of Technology.
Espoo, Finland.
Vesanto, J. (1997). Data Mining Techniques Based
on the Self-Organizing Map. Unpublished master
thesis, Helsinki University of Technology.
Using Self-Organizing Maps for Rural Territorial Typology
Vesanto, J. (1999). SOM based data visualization
methods. Intelligent Data Analysis, 3, 111–126.
doi:10.1016/S1088-467X(99)00013-X
Demartines, P., & Blayo, F. (1992). Kohonen selforganizing maps: is the normalization necessary?
Complex Systems, 6, 105–123.
Vesanto, J., & Alhoniemi, E. (2000). Clustering of the Self-Organizing Map. IEEE Transactions on Neural Networks, 11(3), 586–600.
doi:10.1109/72.846731
Erwin, E., Obermayer, K. K., & Schulter, K.
(1992). Self-Organizing maps: Stationary states,
metastability and convergence rate. Biological
Cybernetics, 67, 35–45. doi:10.1007/BF00201800
Winter, K., & Hewitson, B. (1994). Self organizing
maps - applications to census data. In Hewitson,
B., & Crane, R. (Eds.), Neural nets: applications
in geography. Amsterdam: Kluwer.
Fausett, L. (1994). Fundamentals Neural Networks: Architectures, algorithms, and applications. Prentice Hall.
Wu, S., & Chow, T. W. (2004). Clustering of the
self-organizing map using a clustering validity index based on inter-cluster and intra-cluster density.
Pattern Recognition, 37, 175–188. doi:10.1016/
S0031-3203(03)00237-1
ADDITIONAL READING
Babu, G. P. (1997). Self-organizing neural networks for spatial data. Pattern Recognition Letters,
18, 133–142. doi:10.1016/S0167-8655(97)000032
Bishop, C. (1995). Neural Networks for Pattern
Recognition. Oxford University Press.
Cai, Y. (1995). Artificial neural-network method
for soil-erosion forecasting. Bodenkultur, 46(1),
19–24.
Cereghino, R., Giraudel, J., & Compin, A. (2001).
Spatial analysis of stream invertebrates distribution in the Adour-Garonne drainage basin (France),
using Kohonen self organizing maps. Ecological
Modelling, 146, 1–3, 167–180. doi:10.1016/
S0304-3800(01)00304-0
Couclelis, H. (1998). Geocomputation in context.
In Longley, P., Brooks, S., McDonnell, R., &
Macmillan, B. (Eds.), Geocomputation: a primer.
John Wiley and Sons.
Fischer, M., & Getis, A. (1996). Recent developments in spatial analysis. Springer.
Foody, G. (1999). Applications of the self-organising feature map neural network in community
data analysis. Ecological Modelling, 120, 97–107.
doi:10.1016/S0304-3800(99)00094-0
Gahegan, M. (1999). What is Geocomputation? Transactions in GIS, 3, 203–206.
doi:10.1111/1467-9671.00017
Gahegan, M., Takatsuka, M., Wheeler, M., &
Hardisty, H. (2002). Introducing GeoVISTA
Studio: an integrated suite of visualization and
computational methods for exploration and
knowledge construction in geography. Computers,
Environment and Urban Systems, 26, 267–292.
doi:10.1016/S0198-9715(01)00046-1
Haese, K. (1998). Self-Organizing Feature Maps
with Self-Adjusting Learning Parameters. IEEE
Transactions on Neural Networks, 9, 1270–1278.
doi:10.1109/72.728376
Hewitson, B., & Crane, R. (1994). Neural nets:
applications in geography. Kluwer.
Ji, C. Y. (2000). Land-use classification of remotely
sensed data using self-organizing feature mapa
neural networks. Photogrammetric Engineering
and Remote Sensing, 66(12), 1451–1460.
125
Using Self-Organizing Maps for Rural Territorial Typology
Kaski, S., Kangas, J., & Kohonen, T. (1998).
Bibliography of Self-Organizing Map (SOM)
Papers: 1981–1997. Neural Computing Surveys,
1, 102–350.
Koga, D. (2003). Medidas de Cidades: entre
territórios de vida e territórios vividos. Cortez.
Kohonen, T. (1990). The Self-organizing maps.
Proceedings of the IEEE, 78(9), 1464–1480.
doi:10.1109/5.58325
Kropp, J. (1998). A neural network approach to
the analysis of city systems. Applied Geography
(Sevenoaks, England), 18(1), 83–96. doi:10.1016/
S0143-6228(97)00048-9
Lo, Z., Yu, Y., & Bavarian, B. (1993). Analysis of
the convergence properties of topology preserving
neural networks. IEEE Transactions on Neural
Networks, 4, 207–220. doi:10.1109/72.207609
Muñoz, A., & Muruzábal, J. (1998). Self-organizing maps for outlier detection. Neurocomputing,
18, 33–60. doi:10.1016/S0925-2312(97)00068-4
Oja, M., Kaski, S., & Kohonen, T. (2003). Bibliography of Self-Organizing Map (SOM) Papers:
1998-2001 Addendum. Neural Computing Surveys, 3, 1–156.
Openshaw, S., & Abrahart, S. (2000). Geocomputation. Taylor & Francis.
Openshaw, S., & Openshaw, C. (1997). Artificial
intelligence in geography. John Wiley.
Purvis, M., Zhou, Q., Cranefield, S., Ward, R.,
Raykov, R., & Jessberger, D. (2001). Spatial information modelling and analysis in a distributed environment. Ecological Modelling & Software, 16,
439–445. doi:10.1016/S1364-8152(01)00014-7
126
KEY TERMS AND DEFINITIONS
Artificial Neural Network (ANN): ANN is
a computational model that simulates functional
aspects of biological neural networks. It consists
of an interconnected set of artificial neurons and
processes information using a parallel approach
to computation. ANN is usually used to find patterns in datasets.
Best Match Unit (BMU): BMU is the neuron
j closer to some input xk than other neurons. It can
also be called winner neuron.
Codevectors: Codevectors are the set of wj,
j=1,2,…,m, m is the number of neurons.
Learning Algorithm: Algorithm that allows
computers to evolve behaviors based on empirical data, such as from sensor data or databases.
Neural Topology: Neural topology refers to
the layout of connected neurons.
Quantization: Quantization is the process of
mapping a large set of possible discrete values
(or a continuous range of values) by a relatively
small and finite set of discrete values which is
also on that original interval.
Voronoi Region: Voronoi region is a kind of
a metric space determined by Euclidean distances
to a specified discrete set of objects (codevectors
wj). Each Voronoi region is formed by all input
vectors xk closer to each reference object (wj) than
any other codevector.
ENDNOTE
1
Municipalities that concentrate services and
opportunities for the rest of the same Territory of Identity.